From agent-teams
Create a NEW agent-teams initiative and hand it to a background DRI. Use when asked to "start a new initiative in the background", "kick off / dispatch a background DRI", "spin off separable work as its own initiative", or when invoked as /dri-dispatch <problem statement>. Creates a dedicated worktree, registers the initiative, and launches a hands-off background DRI session that drives it to a PR. Does NOT make THIS session the DRI — use /dri for that.
How this skill is triggered — by the user, by Claude, or both
Slash command
/agent-teams:dri-dispatchThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
You dispatch a new initiative; you do not become its DRI. This skill sets up a fresh, isolated checkout, registers the initiative in the global workspace, and launches a **background** `/dri` session that owns it end-to-end. The current session stays free.
You dispatch a new initiative; you do not become its DRI. This skill sets up a fresh, isolated checkout, registers the initiative in the global workspace, and launches a background /dri session that owns it end-to-end. The current session stays free.
Use this when:
For becoming the DRI in this session, use /agent-teams:dri instead.
ateam toolateam is on PATH — it ships as a prebuilt binary in the plugin's bin/ (auto-added to PATH; installed/verified by /setup-agent-teams). Call it as bare ateam everywhere this document shows ateam. One allowlist entry covers all subcommands: Bash(ateam:*).
🚨 CARDINAL RULE. The GLOBAL workspace (reached ONLY via ateam) holds ONLY initiative-tracking beads and role memories. Registering an initiative is the ONE write you make there, and ateam register is the only sanctioned way to do it — NEVER bd -C <global> create. All work beads (the planner's decomposition, feature/task/discovery beads) live in the PROJECT repo and are created by the background DRI and its team, not here.
ateam is on PATH: run ateam ws. If it errors or is not found, tell the human to run /setup-agent-teams and stop.ateam audit. It must report clean before you add anything.The LLM's job here is judgment only — everything mechanical is handled by ateam dispatch in step 3.
ateam show <id>. A bare one-liner starves its clarify phase; a rich context block lets it proceed with the right assumptions. Write the context to a temp file (e.g. /tmp/ateam-ctx-<slug>.txt) using the Write tool. If there is genuinely no additional context, skip the file and omit --body-file.--repo <abs-path> ONLY when you are not confident: cwd is not inside any repo, the problem clearly refers to a different project you cannot locate, or more than one repo plausibly fits.--base-branch <b> only when the human implies a non-default base or there is genuine ambiguity — a wrong base is expensive to unwind.Run a single call. Everything deterministic (slugify, git worktree add, initiative register, background DRI launch) is handled inside ateam dispatch:
ateam dispatch --problem "<one-line problem statement>" --body-file <tmpfile> [--repo <abs-path>] [--base-branch <branch>]
--problem is the one-line title. --body-file carries the full context block you wrote in step 2 (schema lines come first automatically; the context is appended after them). Omit --body-file only when there is truly no additional context to pass.
dispatch fail-fasts (non-zero exit) on: not-a-git-repo, empty slug, worktree-slug collision, or a --body-file path that cannot be read. It never prompts. On success it prints:
initiative_id: <id>
worktree: <abs-path>
slug: <slug>
base_branch: <branch>
🚨 CARDINAL RULE. ateam dispatch performs the ONE write to the global workspace (initiative registration via the same register path). All work beads (planner's decomposition, feature/task/discovery beads) live in the PROJECT repo and are created by the background DRI and its team, not here.
Relay the output dispatch printed. Tell the human:
claude agents # list background sessions
claude logs <session-id> # recent output without attaching
claude attach <session-id> # open it in this terminal
claude stop <session-id> # abort early OR reap a finished idle session
When the background DRI finishes, it ends its turn and the session stays idle — it does NOT self-stop. It appears as idle in claude agents; use claude stop <session-id> to stop it when you are done with it.
Any human gate the background DRI parks on surfaces through ateam human-list and the /initiatives dashboard — so a needed decision is discoverable without tailing logs.
A backgrounded DRI has no human attached to answer prompts, so it runs with --permission-mode bypassPermissions (what ateam dispatch sets when launching the background DRI). This requires a one-time interactive acceptance of bypass mode on the machine first. A bypass-mode DRI edits a real repo unattended — only dispatch one for well-scoped work, and confirm with the human first when the initiative touches sensitive tooling or infrastructure.
Reference: https://code.claude.com/docs/en/agent-view
npx claudepluginhub mgt-insurance/agent-teamsCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.